Harvey

Contribution to AI

The distance between a model that writes plausible legal prose and one a partner will put their name to is where this company did its work. Allen & Overy's decision in February 2023 to put the tool in front of roughly 3,500 lawyers across 43 offices — following a beta that had already absorbed some 40,000 queries — was the first firmwide deployment of a GPT-based system at a firm of that size, and PwC's alliance a month later extended the same bet to the legal arm of a Big Four accountancy. What made those deployments interesting was not the drafting. It was that hallucination in a legal filing is a professional-conduct problem rather than an inconvenience, so the engineering had to concentrate on grounding, provenance and verifiable citation rather than fluency. The evaluation work followed from that. BigLaw Bench was assembled from real billable time entries and scored against rubrics measuring how much of a lawyer's actual work product a system completed and whether its references held up — a deliberate rejection of bar-exam-style multiple choice as evidence of competence. Later integration of primary law and Shepard's Citations conceded the same point: a general model needs an authority layer beneath it. Commercially the demonstration was that a vertical application company could charge enterprise prices and build defensible depth on top of somebody else's foundation model.

Drafted with AI and edited by hand (claude-opus-5, reviewed 2026-08).